BioCosmos: computer vision and large language models for large-scale image-based biodiversity data and trait exploration

Heru Handika , Moritz D. Lürig, Kira L. Nichtawitz, Oliver M. Dobon, Venkata S. Yarava, Justin K. Sobayo, Breanna Shi, Michael J. Elliott, Jose A. B. Fortes, Arthur Porto

Millions of natural history specimen images are now online, but poor discoverability keeps their rich phenotypic and evolutionary data out of reach for most users. We developed BioCosmos, a web platform that combines conventional biodiversity databases with emerging computer vision and large language models for image retrieval, visual feature recognition, and natural language queries. Its first implementation, LepiVerse (https://lepiverse.com), is a butterfly web portal with more than 600,000 specimen images aggregated from natural history museums worldwide via GBIF, SCAN, and Ecdysis, and curated through taxonomic harmonization and georeference validation pipelines. BioCosmos features visual similarity, interactive morphospace plots, and text- and image-based searches, alongside species pages enriched with data from GenBank, LepTraits, Wikipedia, and Crossref. The platform demonstrates how an open-source, low-cost emerging AI can make vast image collections accessible to a broad audience, while highlighting the digitization gap between the Global North and South.